Difference between revisions of "Document:MIT-MFin-optional-2017"

From LQ's wiki
Jump to: navigation, search
 
(3 intermediate revisions by one user not shown)
Line 4: Line 4:
  
 
{{CVContent|
 
{{CVContent|
I am applying to the 18 month pilot for two reasons:
+
I am applying to the 18 month programme for two reasons:
 
}}
 
}}
  
Line 11: Line 11:
 
}}
 
}}
 
{{CVContent|
 
{{CVContent|
Since most of my professional and academic experience was in technology, deferring recruitment to the returning fall term allows me to concentrate academically throughout the first academic year, to make my case more competitive. Most importantly, the three months of internship would have very high impact in terms of working experience in finance.
+
Since most of my professional and academic experience was in technology, deferring recruitment to the returning fall term allows me to concentrate academically throughout the first year, to make my case more competitive. Most importantly, the three months of internship would have very high impact in terms of working experience in finance.
 
}}
 
}}
 
{{CVUSSubheader
 
{{CVUSSubheader
Line 17: Line 17:
 
}}
 
}}
 
{{CVContent|
 
{{CVContent|
The 18-format allows more time to work towards the thesis, or to take additional units to inform it.
+
The 18-month format gives me more time to either take the thesis as far as I can, or to take additional units to inform it.
  
 
My research is about the application of deep convolutional models to financial time series analysis. Such methods summarize time series (potentially with multiple channels, e.g. trade and market depth data) in terms of nonlinear functions of convolutions with filters learned automatically through optimization. The motivation is to capture relationships missed by hand-engineered or low-level features (e.g. technical indicators, RSI, correlation etc.) which are not necessarily optimal -- [Foundations of Technical Analysis. Lo 2000: section V]
 
My research is about the application of deep convolutional models to financial time series analysis. Such methods summarize time series (potentially with multiple channels, e.g. trade and market depth data) in terms of nonlinear functions of convolutions with filters learned automatically through optimization. The motivation is to capture relationships missed by hand-engineered or low-level features (e.g. technical indicators, RSI, correlation etc.) which are not necessarily optimal -- [Foundations of Technical Analysis. Lo 2000: section V]

Latest revision as of 10:36, 5 January 2017


I am applying to the 18 month programme for two reasons:

1. Timing and marginal utility of the internship

Since most of my professional and academic experience was in technology, deferring recruitment to the returning fall term allows me to concentrate academically throughout the first year, to make my case more competitive. Most importantly, the three months of internship would have very high impact in terms of working experience in finance.

2. Research and thesis

The 18-month format gives me more time to either take the thesis as far as I can, or to take additional units to inform it.

My research is about the application of deep convolutional models to financial time series analysis. Such methods summarize time series (potentially with multiple channels, e.g. trade and market depth data) in terms of nonlinear functions of convolutions with filters learned automatically through optimization. The motivation is to capture relationships missed by hand-engineered or low-level features (e.g. technical indicators, RSI, correlation etc.) which are not necessarily optimal -- [Foundations of Technical Analysis. Lo 2000: section V]

Such features from multiple time series could be used to, for example, forecast and refine the next-period expected return in portfolio management, or used in a manner that resembles pairs trading. Moreover, the representation lends itself to transfer learning; suppose we have insufficient information at long time-scales. We could train on higher-frequency data and either use the high-frequency features a priori or additionally refine them with what we have. In this way, we potentially accrue an information advantage by learning common signals across multiple time scales, and the gradual transition between long and short term is modelled naturally.


Thank you for your kind consideration.